For years, catching AI models that could generate child sexual abuse material seemed impossible. The law forbids creating such content, even for testing purposes — but that meant nobody could safely check whether a dangerous model had been uploaded online. It was, as one researcher put it, a "huge blind spot."
Now, a team at MIT may have found a way to close that gap. Graduate student Vinith Suriyakumar and associate professors Ashia Wilson and Marzyeh Ghassemi have developed a new auditing technique that can determine whether an AI model has been trained to produce illegal content — without ever generating a single image.
The problem is urgent. The National Center for Missing and Exploited Children received over 1.5 million reports of AI-generated child sexual abuse material in 2025 alone, a staggering jump from just 67,000 in 2024. That twenty-two-fold increase in one year reflects how quickly generative AI tools have proliferated online.
Currently, the standard way to test whether an AI model is dangerous is to prompt it and inspect what comes out. But that method has two major flaws: it doesn't scale, and repeatedly viewing harmful outputs takes a psychological toll on human reviewers. For illegal content, it's also simply not allowed under U.S. law.
The MIT team's approach sidesteps these problems entirely. Instead of looking at what a model outputs, they examine how the model's internal structure changes during a process called fine-tuning — a common technique where someone adapts a pre-existing AI model for a specific purpose. The researchers, working with the child safety nonprofit Thorn, use a method called Gaussian probing to feed random data through the model and analyze how it processes that information internally. They never actually run the model to completion or prompt it to generate anything.
"We never run the model all the way to the end or prompt the model, so we never generate images," Suriyakumar explained.
When tested, the technique identified models that had been specialized to produce illegal content with 100 percent accuracy. A hosting platform could use it to flag dangerous models and remove them before they cause harm.
For Suriyakumar, the significance goes beyond the technical achievement. "Before, we had no way of measuring this," he said. "It was a huge blind spot that some people were taking advantage of. Now, we can address an AI safety problem that is having severe negative impacts."
The research was presented at the International Conference on Machine Learning and involved collaboration with colleagues at Boston University and Thorn. The team hopes this work opens a new pathway for both AI platforms and law enforcement to detect dangerous models — and keep children safer in an era of rapidly spreading AI technology.